move tty_kref_put() outside of __cleanup_signal()
Subsystems (3)
PATCH 1/1 move tty_kref_put() outside of __cleanup_signal()
tty_kref_put() have two callsites in copy_process() pathes,
1. if copy_process() suceeds it is called before we copy signal->tty from parent
2. otherwise it is called from __cleanup_signal() under
bad_fork_cleanup_signal: label
In both cases tty_kref_put() is not right and unneeded because we don't have the balancing tty_kref_get(). Fortunately, this is harmless because this can only happen without CLONE_THREAD, and in this case signal->tty must be NULL.
Remove tty_kref_put() from copy_process() and __cleanup_signal(), and change another caller of __cleanup_signal(), __exit_signal(), to call tty_kref_put() by hand.
I hope this change makes sense by itself, but it is also needed to make ->signal refcountable.
---
kernel/fork.c | 2 -- kernel/exit.c | 1 + 2 files changed, 1 insertion(+), 2 deletions(-)
kernel/fork.c renamed from kernel/fork.c~7_TTY_PUT +0 -2
@@ -896,7 +896,6 @@ static int copy_signal(unsigned long clovoid __cleanup_signal (struct signal_struct *sig) {thread_group_cputime_free (sig); -tty_kref_put (sig->tty);kmem_cache_free (signal_cachep, sig); }
@@ -1264,7 +1263,6 @@ static struct task_struct *copy_process( p->nsproxy->pid_ns->child_reaper = p; p->signal->leader_pid = pid; -tty_kref_put (p->signal->tty); p->signal->tty =tty_kref_get (current->signal->tty);attach_pid (p,PIDTYPE_PGID ,task_pgrp (current));attach_pid (p,PIDTYPE_SID ,task_session (current));
kernel/exit.c renamed from kernel/exit.c~7_TTY_PUT +1 -0
@@ -150,6 +150,7 @@ static void __exit_signal(struct task_st* see account_group_exec_runtime(). */ task_rq_unlock_wait (tsk); +tty_kref_put (sig->tty);__cleanup_signal (sig); } }
kernel/fork.c static struct task_struct *copy_process( · line 1266
Quoted 2 lines
kernel/fork.c renamed from kernel/fork.c~7_TTY_PUT +0
-1-tty_kref_put (p->signal->tty); p->signal->tty =tty_kref_get (current->signal->tty);
That bit needs commenting clearly or a WARN_ON() that p->signal->tty is NULL before the get otherwise when the assumption is broken the flaw will be subtle and hard to find.
kernel/exit.c static void __exit_signal(struct task_st · line 153
Quoted 9 lines
kernel/fork.c renamed from kernel/fork.c~7_TTY_PUT +0
-0attach_pid (p,PIDTYPE_PGID ,task_pgrp (current));attach_pid (p,PIDTYPE_SID ,task_session (current));kernel/exit.c renamed from kernel/exit.c~7_TTY_PUT +1
-0@@ -150,6 +150,7 @@ static void __exit_signal(struct task_st* see account_group_exec_runtime(). */ task_rq_unlock_wait (tsk); +tty_kref_put (sig->tty);
and a sig->tty = NULL assignment to trap races might not go amiss here perhaps ?
Quoted 7 lines · Alan Cox
kernel/fork.c static struct task_struct *copy_process( · line 1266
Quoted 3 lines
kernel/fork.c renamed from kernel/fork.c~7_TTY_PUT +0
-1-tty_kref_put (p->signal->tty); p->signal->tty =tty_kref_get (current->signal->tty);That bit needs commenting clearly or a
WARN_ON()thatp->signal->ttyis NULL before the get otherwise when the assumption is broken the flaw will be subtle and hard to find.
Well. It must be NULL. This tty_kref_get() was added by 9c9f4ded90a59eee84e15f5fd38c03d60184e112, and the same commit also added sig->tty = NULL into copy_signal(). With the recent changes in copy_signal() it must be NULL due to zalloc().
But this doesn't matter. We should not do tty_kref_put() even if it is not NULL. If we do "put", we need the balancing "get" which we do not have. That is why this "get" is not only unneeded, but wrong imho.
Probably the little comment makes sense, but I don't really understand what it should explain. To me, this is like list_add_tail(&init_task.tasks) few lines below. We do not have the comment to explain the "missing" list_del(). This task is new, nobody can see/use it before we drop the locks. NULL or not, its signal->tty is just uninitialized yet.
Quoted 10 lines
kernel/exit.c static void __exit_signal(struct task_st · line 153
Quoted 7 lines
kernel/exit.c renamed from kernel/exit.c~7_TTY_PUT +1
-0@@ -150,6 +150,7 @@ static void __exit_signal(struct task_st* see account_group_exec_runtime(). */ task_rq_unlock_wait (tsk); +tty_kref_put (sig->tty);and a
sig->tty= NULL assignment to trap races might not go amiss here perhaps ?
Indeed ;)
The subsequent patches will do this, we need more changes anyway. Currently this doesn't matter because we are going to kfree() this memory unconditionally. But when we pin ->signal to task_struct, we should clear ->signal->tty before we drop ->siglock, then tty_kref_put().
Oleg.
Quoted 2 lines
This task is new, nobody can see/use it before we drop the locks. NULL or not, its
signal->ttyis just uninitialized yet.
Fair enough.
Quoted 12 lines
Quoted 11 lines
kernel/exit.c static void __exit_signal(struct task_st · line 153
Quoted 8 lines
kernel/exit.c renamed from kernel/exit.c~7_TTY_PUT +1
-0@@ -150,6 +150,7 @@ static void __exit_signal(struct task_st* see account_group_exec_runtime(). */ task_rq_unlock_wait (tsk); +tty_kref_put (sig->tty);and a
sig->tty= NULL assignment to trap races might not go amiss here perhaps ?
Quoted 6 lines
Indeed ;)
The subsequent patches will do this, we need more changes anyway. Currently this doesn't matter because we are going to
kfree()this memory unconditionally. But when we pin ->signal totask_struct, we should clear ->signal->tty before we drop ->siglock, thentty_kref_put().
Ok - yes the moment you start refcounting ->signal that changes (or do you expect to free ->tty when you destruct the signals ?)
Quoted 14 lines · Alan Cox
Quoted 14 lines
Quoted 11 lines
kernel/exit.c static void __exit_signal(struct task_st · line 153
Quoted 8 lines
kernel/exit.c renamed from kernel/exit.c~7_TTY_PUT +1
-0@@ -150,6 +150,7 @@ static void __exit_signal(struct task_st* see account_group_exec_runtime(). */ task_rq_unlock_wait (tsk); +tty_kref_put (sig->tty);and a
sig->tty= NULL assignment to trap races might not go amiss here perhaps ?Indeed ;)
Quoted 7 lines
Quoted 4 lines
The subsequent patches will do this, we need more changes anyway. Currently this doesn't matter because we are going to
kfree()this memory unconditionally. But when we pin ->signal totask_struct, we should clear ->signal->tty before we drop ->siglock, thentty_kref_put().Ok - yes the moment you start refcounting ->signal that changes (or do you expect to free ->tty when you destruct the signals ?)
I think signal->tty should be freed (and nullified under ->siglock) when the last thread exits.
The goal is to make ->signal immutable, so that it would be always safe to dereference task->signal if you have a reference to task_struct. But I don't think get_task_struct() should defer tty_kref_put(), and besides put_task_struct() must be safe in any context.
See also http://marc.info/?l=linux-kernel&m=126885423426183
Oleg.
Thanks,
Roland